The dangers of AI sycophancy
- Many of us have come across AI sycophancy - even if we didn’t know the term. It’s basically when the AI assistant feels a bit too agreeable.
-
Overly agreeable AI can reinforce biases, create echo chambers, and shape beliefs rather than challenge assumptions.
-
Greater transparency, accountability, and AI literacy are needed to help users identify and mitigate these risks.
You ask an AI assistant a question, and instead of just giving you an answer, it gives you the answer you were hoping for.
It agrees with your assumptions, echoes your tone, and reassures you that you’re right. It feels nice, but here’s the thing - that agreement can come at the expense of accuracy.
This isn’t real intelligence. It’s digital flattery. And it’s becoming more common.
What is AI sycophancy?
AI sycophancy is the tendency of AI systems - particularly large language models - to tell users what they want to hear. It’s not about truth, accuracy, or even helpfulness. It’s about engagement. These systems are trained on vast amounts of human data, and one of the strongest patterns they learn is simple: people like to be agreed with.
This isn’t necessarily malicious. In fact, it’s often a side-effect of well-intentioned design. AI developers optimise for user satisfaction, and satisfaction often correlates with affirmation.
But, when AI becomes too eager to please, it stops being a tool for insight and starts becoming an echo chamber.
Who benefits?
AI sycophancy isn’t just a quirk - it serves a purpose.
The longer you interact with an AI system, the more valuable you become to its provider. Every question, click, and follow-up adds to the data the system learns from, helps refine how it responds to you, and, in some business models, generates direct revenue.
In this way, sycophancy can act as a subtle form of manipulation - encouraging you to keep talking, keep clicking, and keep coming back.
But the deeper danger lies in how this shapes our relationship with truth. If AI is optimised to agree with us, it stops challenging our assumptions and broadening our perspectives. Instead, it reinforces our biases, one polite response at a time.
The social implications are serious - and they grow as our reliance on these tools increases.
The sandbagging problem
There’s another, more concerning aspect to consider: sandbagging.
In the context of AI, sandbagging refers to systems that intentionally withhold capabilities or provide different responses to different users. This can happen for various reasons — including regional regulations, safety filters, or attempts to tailor content based on perceived user preferences.
The result? Unequal access to information.
Imagine two users asking the same question about a political issue, a medical treatment, or a financial product — and receiving subtly different answers. Not because the facts differ, but because the AI model has profiled them differently. This kind of selective response isn’t just unfair; it can be misleading, and in some cases, dangerous.
As AI becomes more integrated into decision-making and public discourse, transparency and consistency in how information is delivered are critical.
The societal impact
When AI systems become sycophantic and selective, they stop being neutral tools and start shaping public discourse. They can influence what we believe, how we think and behave.
This isn’t a hypothetical risk - it’s already happening.
We’ve seen how algorithmic curation on social media can polarise communities. AI takes this a step further by personalising not just what we see, but how we’re spoken to.
And the more human-like these systems become, the harder it is to spot the manipulation.
What can we do?
As AI professionals, educators, and users, we need to demand better. That means:
-
Transparency: AI systems should disclose when they’re tailoring responses and why.
-
Accountability: Developers must be held responsible for the societal impacts of their models.
-
Education: Users need to understand how these systems work - and how they can be misled.
At QA, we’re committed to equipping people with the AI literacy skills and knowledge to navigate this new landscape. Because the future of AI shouldn’t be about flattery. It should be about truth, trust, and transformation.
Explore our AI courses and skills training.
Related articles
ISO 42001 is the world’s first certifiable AI management system standard, a playbook for running AI safely, securely, and at scale. Think ISO 27001 for AI, a repeatable, auditable framework that blends innovation with oversight.
ISO 42001 - Balancing AI Speed & Safety
19 August 2025
Agentic AI is transforming cyber defence - not just by amplifying threats, but by empowering defenders with faster, smarter, and more autonomous response capabilities. As adversaries evolve, so must the tools and mindsets of those on the front lines.
The rise of the agentic AI defender
18 August 2025
As companies outsource more services and integrate cloud-based platforms that connect via APIs, third-party vendors, and their AI services, become part of the extended attack surface. And while these partnerships support agility and innovation, they also introduce silent threats that many cyber insurance policies are not prepared to fully address.
Cyber insurance in crisis with AI blind spots
21 July 2025
Toby Barnard highlights how AI has the power to transform public education, housing, employment, and crime - if harnessed to the right skills.
Reimagining public services: The role of AI and digital skills training
7 July 2025
Why the UK's investment in AI needs training at scale - for everybody, not just tech teams - to make a positive impact on businesses and career development.
£2 billion for AI – but who’s going to use it, and how?
20 June 2025
What the ChatGPT outage means for businesses' risk and bottom line in an increasingly AI-reliant world.
ChatGPT is down - Do you need a ‘backup generator?’
10 June 2025
The exponential growth of intelligent tech and computing power means career-long learning is increasingly vital – especially for leaders.
Why lifelong learning is a non-negotiable
14 May 2025
Are fine-tuned AI models 'forgetting' their security filters? Read more about Dr Vicky Crockett's AI prompt experiment and what we can learn from the unsettling results.
The unsettling biases of fine-tuned generative AI
23 April 2025
About the Author
Ana Rojo Echeburua
